14655

13200

كارشناسي ارشد

سيستم هاي اقتصادي-اجتماغي

اصفهان : دانشگاه صنعتي اصفهان

1398

يازده، 71ص، مصور، جدول، نمودار

مهدي بيجاري

مهدي خاشعي

علي زينلي همداني، رضا حجازي

1398/03/11

كتابنامه

مهندسي صنايع

مهندسي صنايع و سيستم ها

1398/03/11

25036063

71 Application of artificial neural networks for forecasting the leakage of gas polyethylene pipes Author Saba Tamizi s tamizi@in iut ac ir Department Industrial and Systems EngineeringDegree MSc Language PersionSupervisor Mehdi Bijari bijari@cc iut ac irAdvisor Mehdi Khashei khashei@cc iut ac irAbstractFor half a century in gas distribution networks polyethylene pipes with advantages such asnon corrosion ease of implementation and operation cost reduction considered as areplacement for metal pipes Considering the Irrecoverable gas leakage events leakage forecast has great importanceand this forecast can be used as a strategy for designing gas pipeline inspection programs Because of the lack of research in the field of the forecast of polyethylene pipes leakage inthis thesis the leakage degree will be forecasted by collection data proportional to theeffective factors leakage in this type of pipes For this purpose 32 factors were identifiedas the primary factors affecting the leakage of these pipes Finally by examining the dataand opinion of the experts eight factors were selected as the final factors First the classic logistic regression model was used to forecast which is a well knownlinear statistical method for classification Since the accuracy of the forecast is one of themost important factors in the choice of forecast method and also in the world of leakagedata in addition to linear patterns there are nonlinear patterns it was decided that nonlinearmodels should also be used to improve accuracy For this reason the soft computing tool which includes two models of the artificial neural network the multilayer perceptron neuralnetwork MLP and radial base network RBF as well as a support of the vector machine SVM were used For each of these models different architectures were considered toidentify the best structure of each model Finally the accuracy of the best structure of eachmodel was compared with each other and it was concluded that multilayer perceptronnetworks are the best single model to forecast the degree of leakage of the collected data To improve the classification accuracy a hybrid model was used to simulate linear andnonlinear patterns simultaneously The best single model MLP with an accuracy of88 80 a sensitivity of 88 19 and MSE 0 111 and a hybrid model with an accuracy of88 58 sensitivity of 89 86 and MSE of 0 080 degrees of leakage forecasted Finally comparing the results as expected the hybrid model was more efficient to forecast theleakage degree than the single model with respect to performance evaluation criteria Key word Leakage polyethylene pipe classification logistic regression artificial neuralnetwork support vector machine hybrid model

مهدي بيجاري

مهدي خاشعي

علي زينلي همداني، رضا حجازي